Webinar Recording & Key Takeaways
Full Webinar Recording (1 hour)
Slides
Key Highlights
- Overview of AI agents - defined as LLMs with a mission, autonomy, and tools to carry out specific tasks
- Demonstration of how to create and configure agents for practical use cases (invoice processing, news digestion)
- Examples of both self-service agents and custom-built agents for specific business needs
- Discussion of agent design patterns - including agent teams, pipeline approaches, and collaborative models
- Practical tips for implementing agents safely, starting with simple use cases before moving to complex ones
- Benefits of using AI agents: time savings (95% reduction in manual work in one case study), improved quality, and better use of human resources
- Options for implementing agents: building from scratch, using open-source platforms, or using SaaS platforms
- Live demonstration of creating agents that can use various tools like email, Slack, document creation, web scraping
- Importance of human-agent collaboration and defining clear boundaries for agents
Demo 1 - Good News Agent
The MrChill demo showcased a simple AI agent that transforms positive news into bedtime stories.
With just a few instructions, the agent scanned a Swedish news site, created an original story based on positive news, generated a children's book-style illustration, and narrated it with a Scottish accent. The demo concluded by setting MrChill to automatically perform this task every evening at 8:00 PM and email the results—demonstrating how easily complex workflows can be automated with AI agents.
Demo 2 - Invoice Audit and Routing
The Freddy demo showcased an invoice processing agent that validates and routes incoming invoices automatically. After receiving company guidelines and department contact information, the agent analyzed invoices, flagged suspicious ones via Slack, and forwarded them to the appropriate teams. The agent also maintained a log of all processed invoices for tracking purposes. This demonstration illustrated how AI agents can efficiently automate routine business workflows while maintaining transparency in their decision-making process.
Upcoming AI Agent Training
Two-Day Intensive Course
Online in english, https://www.ymnig.ai/courses/autonomous-ai-agents
In-person, Stockholm, Sweden, https://www.ymnig.ai/courses/ai-agenter
Transform your organization with practical AI agent knowledge:
- From concept to implementation in 48 hours
- Hands-on experience with leading
Ready to Start Your AI Agent Journey?
Book an Agent Discovery Meeting
Let's explore how AI agents can benefit your specific business needs.
https://www.ymnig.ai/book-a-discovery-call
Stay Connected
Follow us for the latest updates on AI agent developments and upcoming events:
FAQ during the Seminar
About the Platform and Technology
Which LLMs do you use?
We currently use Claude 3.7 Sonnet for everything. Our platform supports DeepSeek, Llama, ChatGPT, and Grok, but Claude has been the best default for most use cases. As of today (2025-04-11), we are currently evaluating Gemini 2.5. We can configure the platform to use other models if customers prefer their own internal solutions.
Are you using an Agent framework like Langchain or LlamaIndex?
No. These frameworks were getting in the way, so we build directly on the underlying APIs. This gives us more control and flexibility, and we can move faster.
Are you planning to use MCP (Model Context Protocol) for tool usage?
We are looking into it, among other ways to integrate with the world.
Can your platform integrate with external tools?
Yes. Our agents have an API with endpoints that allow your systems to communicate with them. All agents can call any HTTP endpoint, so if you have a tool that you host, we can use it. We also add new tools on demand - it typically takes about 5-10 minutes to implement the simpler ones.
What have you learned about web scraping?
We currently use a third-party service called Firecrawl that works perfectly for our needs. This has eliminated previous bottlenecks with web scraping.
Can the platform generate Word documents from text?
While not currently implemented, adding this capability would take approximately 5 minutes to develop. Any new capabilities requested by users can be quickly added to the platform.
Can I use local LLMs with your platform? Can the platform run locally?
The platform currently uses cloud-based LLMs. We don't currently offer a fully local deployment option, but we're exploring customization options based on customer needs. We could easily integrate with a locally hosted LLM, but so far, we have not found an LLM of good enough quality to run locally. The agents would become less intelligent with current local models.
Getting Started and Usage
What's needed to start using your platform?
Currently, we're running a private beta. You can sign up through the links provided after the seminar. The demo you saw serves as your onboarding - the platform is designed to be intuitive for self-service.
Is there a lock-in with your platform?
No. All platforms use similar principles with instructions and prompts. Moving to another platform simply requires taking your prompts and leaving - there's no proprietary lock-in.
How does your platform compare to just using available APIs?
Our platform provides a conversational interface for creating agents instead of requiring coding. This makes it more accessible while still providing the capabilities available through direct API usage.
Are agent behaviors deterministic?
No. Whenever you use an LLM, intelligence is involved which means giving up some predictability. The more complex the task, the less predictable the behavior. You can increase reliability with good models (like Claude 3.7 Sonnet) and limiting the scope of tools, but 100% determinism requires writing code instead of using LLMs.
How can I test the reliability of agents before handing them off to clients?
We recommend a phased approach to testing reliability. Start with simple tasks and limited tools, then gradually expand capabilities as confidence grows. The platform includes monitoring capabilities to help track agent performance.
Have you used agentic coding for programming your agents?
Yes, for certain aspects like evaluating instructions for safety and scheduling recurring tasks. We use multiple layers of agentic evaluations internally for complex scenarios. Our goal is to build this intelligence into the platform so customers don't need to worry about implementing these layers themselves.
Costs and Business Model
Who pays for the tokens?
As a beta tester, you'll receive several weeks of free tokens.
What about cost management when using models like Claude?
Cost management is built into the platform. For complex flows with large token exchanges, we've implemented optimizations to maintain a good signal-to-noise ratio and manage costs effectively.
Will the platform be available for individual licensing or only for enterprises?
We can discuss various licensing options based on your needs. While the platform isn't open source, we're open to different licensing arrangements.
Security and Privacy
How do you handle intellectual property concerns?
We're aware that feeding company information to AI models raises potential intellectual property concerns. Our platform is designed with security in mind, and we recommend starting with less sensitive data as you become comfortable with the technology.
What is your data privacy policy?
We maintain strict data privacy protocols. Details about our specific policies are provided during the onboarding process.
Learning Resources
Are there books available on this topic?
Yes, there are books available on AI agents and related topics. Henrik Kniberg has written a book, and his colleague Nils has recently completed one as well. We also have books by Johan Falk. All books are available on LeanPub, and you can find links to all of them in our library at https://www.ymnig.ai/library.
What was your biggest "aha moment" working with agents?
Henrik mentioned that his biggest realization was how powerful agents are when you find the right use case, and how human-like they can be. He noted that treating agents as humans you collaborate with tends to work better, even though we know they aren't human. This approach leads to better results when writing prompts.
